Case study 6: The artificial neural network model for

Transcription

Case study 6: The artificial neural network model for
Joint Cefic LRI/Cosmetics Europe/EPAA workshop
Alternatives for Skin Sensitization testing
23-24 April 2015, ECHA offices, Helsinki Finland
Case study 6:
The artificial neural network model
for predicting LLNA EC3
Takao Ashikaga, Shiseido, Yokohama, Japan
Aim of the study
Due to regulatory constrains and ethical concerns, alternatives to
animal testing are needed to predict skin sensitizing potential of
chemicals.
To do risk assessment, potency evaluation is essential and the
relative potency of targeted chemicals has been mainly valuated
by calculating EC3 value, which is obtained from LLNA.
In this study, I propose a skin sensitization potency prediction
model using artificial neural network analysis of data from
multiple in vitro assays. EC3 value can be predicted by using this
model, and the predicted EC3 value can be applied for prediction
of a safe level of human exposure using a Quantitative Risk
Assessment (QRA) approach.
Key events of the AOP and corresponding methods
Key event 1: Peptide reactivity (DPRA or SH test)
Key event 2: Keratinocyte response (KeratinoSensTM or ARE assay)
Key event 3: Dendritic cell activation (h-CLAT)
Dataset used in this study
Test method
DPRA
h-CLAT
KeratinoSens
Developer
P&G
Kao/Shiseido
Givaudan
Total 134 chemicals
Data-sharing
P&G
1. All 134 chemicals were used for model development. Then
10-fold cross (inner) validation was conducted.
2. 10 new (un-learned) chemicals were evaluated to test the
model.
Many thanks P&G, Givaudan and Kao for permission to use this dataset.
Log(h-CLAT)
Log(LLNA EC3 (uM/cm2))
R=0.71
P<0.05
N=134
Log(LLNA EC3 (uM/cm2))
Log(DPRA Lys)
R=0.52
P<0.05
N=134
Log(KeratinoSens)
Log(DPRA Cys)
Relationship between in vitro parameters and EC3
R=0.28
P<0.05
N=134
Log(LLNA EC3 (uM/cm2))
R=0.57
P<0.05
N=134
Log(LLNA EC3 (uM/cm2))
Each descriptor derived from DPRA (Cys), DPRA (Lys), h-CLAT and KeratinoSens
correlated with LLNA EC3, significantly. However, each single indicator seems to be not
enough for risk assessment due to variability. How should we fuse them?
Artificial Neural Network (ANN) analysis
axon
synapses
dendrite
In vitro test data
h-CLAT
DPRA-Cys
...
DPRA-Lys
In vivo test data
• EC3 values for LLNA
positive chemicals
• For LLNA negative
chemicals, 101% was
used as EC3 value
(converted to M/cm2 unit).
KeratinoSens
Other biomarkers
Input layer
LLNA EC3
Hidden layer
Output layer
Log(Predicted LLNA EC3 (uM/cm2) )
Correlation between LLNA and the ANN prediction
R= 0.87(N=134)
RMS error=0.57
-
Log(Published LLNA EC3
*Root mean square (RMS) error =
After 10-fold
cross-validation
(uM/cm2))
 ((measured value - predicted value)
2
) / number of data
This ANN model showed a good correlation between predicted
values and actual values.
Category predictive capacity of the ANN model
5
Predicted
5
4
3
2
Negative(2/3)
Extreme
7
1
0
0
0
4
Strong
1
10
4
0
3
LLNA
3
Moderate
0
3
20
9
7
2
Weak
0
0
6
18
6
1
NS
0
0
0
9
30
EC3<0.1% ; Extreme, 0.1%<EC3<1%; Strong, 1%<EC3<10%; Moderate,
EC3>10%; Weak, NS; not-sensitizing in LLNA or Negative in in vitro
tests. Positive/negative identification: 2 out of 3 approach.
For five category classification (extreme, strong, moderate, weak and
non-sensitizer), the accuracy is 63%.
Test the ANN model by newly evaluated chemicals
This work has been conducted as a joint study between Merck (Dr. Thomas
Broschard) and Shiseido.
Actual LLNA data
Estimated results (In vitro and ANN )
Substance
EC3 (%)
Category
EC3 (%)
Category
Item1
-
Negative
Not soluble in
DPRA
-
Item2
28.6
Weak
33.6
Weak
Item3
-
Negative
0.2
Strong
Item4
-
Negative
Not soluble in
DPRA
-
Item5
-
Negative
45.8
Weak
Item6
(GPMT)
Negative
76.6
Negative
Item7
29.4
Weak
56.2
Weak
Item8
47.7
Weak
37.4
Weak
Item9
-
Negative
24.9
Weak
Item10
14.5
Weak
Co-elusion in
DPRA
-
3 of 10 chemicals were out of applicability domain of DPRA.
6 of the 7 chemicals were predicted almost correctly (weak or non-sensitizer).
1 chemical (all three in vitro tests positive) was overestimated with our model.
Main limitations and uncertainties of the approach
 Because these in vitro methods are submerged cell-based assay
or HPLC assay, water-insoluble chemicals are difficult to
evaluate correctly.
 Because of limited metabolic capability of the cell lines and
experimental conditions, pro-haptens (i.e. chemicals requiring
enzymatic activation) and pre-haptens (i.e. chemicals requiring
oxidation) are thought to be out of the applicability domain.
 This approach is currently not designed to predict complex
mixtures.
 Though the artificial neural network is thought to be a useful
tool especially for prediction of complex biological response,
there is a so-called "black box" in the process of the decision.
Furthermore, we need to discuss how we show its validity.
Conclusions and future…
 Each event in the AOP of skin sensitization (protein-binding,
dendritic cell activation and keratinocyte response) might affect
the potency.
 The correlation between actual EC3 value and predicted one was
good. Therefore, our ANN model can contribute to building a
new QRA evaluation system which is using no animals.
 The results of the newly tested chemicals showed both the
usefulness and limitations of this approach.
 We should discuss more how to prove the validity of the ANN
model.